Security boundaries

DataQI enforces strict security boundaries to ensure that agents and AI models can never access data or execute tools that the requesting user is not explicitly authorised to use. The platform operates on a model of dynamic, user-scoped authentication for document retrieval, while maintaining clear boundaries for globally configured tools.

File system connection boundaries

Enterprise file system routes operate strictly under User Identity Delegation (UID). DataQI never promotes data or executes searches using a global service account. Instead, an agent’s visibility and reach are strictly limited to the authenticated user’s natural permissions.

User A’s connection state or data visibility can never be accessed by, or influence the responses for, User B.

MCP tool boundaries

DataQI connects to external business systems via the Model Context Protocol (MCP). Unlike file system connections, MCP servers are configured globally at the environment level using dedicated service accounts or API keys.

This means that when an autonomous agent is granted access to an MCP tool, it executes that tool using the global identity configured for the MCP server. The tool’s reach and capabilities are defined by the permissions of that service account in the external system, rather than the individual user interacting with the agent.

Relational database query boundaries

The SQL query capability executes generated queries using the credentials configured for the selected relational database connection. DataQI applies several application-level precautions, but the permissions assigned to that database account define the capability’s maximum reach.

Database administrators must provide dedicated, least-privileged credentials with read access limited to the required schemas, tables and views. Application-level SQL validation and transaction rollback do not replace database-level authorisation.

Trust partitioning

When agents process your requests, they combine your natural language query, the system’s instructions, and any relevant context (such as a document you are querying).

To defend against prompt injection—where malicious data in a document might attempt to hijack the agent’s instructions—DataQI employs Trust Partitioning. This security boundary structurally separates user-supplied context from the core system instructions. All untrusted context is tagged, escaped, and wrapped in defensive guards before it is passed to the AI model, significantly reducing the risk of unauthorized instruction execution.